Perinatal, Preterm and Paediatric Image Analysis: 7th International Workshop, PIPPI 2022, Held in Conjunction with MICCAI 2022, Singapore, September ... (Lecture Notes in Computer Science) 3031171160, 9783031171161

Citation preview

LNCS 13575

Roxane Licandro · Andrew Melbourne · Esra Abaci Turk · Christopher Macgowan · Jana Hutter (Eds.)

Perinatal, Preterm and Paediatric Image Analysis 7th International Workshop, PIPPI 2022 Held in Conjunction with MICCAI 2022 Singapore, September 18, 2022, Proceedings

Lecture Notes in Computer Science Founding Editors Gerhard Goos Karlsruhe Institute of Technology, Karlsruhe, Germany Juris Hartmanis Cornell University, Ithaca, NY, USA

Editorial Board Members Elisa Bertino Purdue University, West Lafayette, IN, USA Wen Gao Peking University, Beijing, China Bernhard Steffen TU Dortmund University, Dortmund, Germany Moti Yung Columbia University, New York, NY, USA

13575

More information about this series at https://link.springer.com/bookseries/558

Roxane Licandro · Andrew Melbourne · Esra Abaci Turk · Christopher Macgowan · Jana Hutter (Eds.)

Perinatal, Preterm and Paediatric Image Analysis 7th International Workshop, PIPPI 2022 Held in Conjunction with MICCAI 2022 Singapore, September 18, 2022 Proceedings

Editors Roxane Licandro Massachusetts General Hospital & Harvard Medical School Charlestown, MA, USA Medical University of Vienna Vienna, Austria Esra Abaci Turk Children’s Hospital Boston, MA, USA

Andrew Melbourne King’s College London London, UK Christopher Macgowan The Hospital for Sick Children Research Institute Toronto, ON, Canada

Jana Hutter King’s College London London, UK

ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Computer Science ISBN 978-3-031-17116-1 ISBN 978-3-031-17117-8 (eBook) https://doi.org/10.1007/978-3-031-17117-8 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2022, corrected publication 2023 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

The goal of the Perinatal, Preterm and Paediatric Image Analysis (PIPPI) workshop is to provide a focused platform for the discussion and dissemination of advanced imaging techniques applied to young cohorts. The technical program typically consists of one keynote talk from a prominent figure in the community and the presentation of previously unpublished papers. Emphasis is placed on novel methodological approaches to the study of, for instance, volumetric growth, myelination and cortical microstructure, and placental structure and function or the assessment of new technical innovations for planned intervention. Although techniques applied to MR neuroimaging provide a significant number of submissions, we are delighted to receive submissions making use of other modalities or applied to other target organs or regions of interest such as the fetal heart and the placenta. The main objective of PIPPI is to provide a forum for researchers in the MICCAI community to discuss the challenges of image analysis techniques as applied to the preterm, perinatal and paediatric setting which are confounded by the interrelation between the normal developmental trajectory and the influence of pathology. These relationships can be quite diverse when compared to measurements taken in adult populations and exhibit highly dynamic changes affecting both image acquisition and processing requirements. Furthermore, this forum will facilitate the presentation and detailed discussion of novel and speculative works, which may be outside the scope of the main conference but are essential for the advancement of modeling and analysis of medical imaging data. Additionally, discussion of these works within a focused group may initiate new collaborations. The application of sophisticated analysis tools to fetal, neonatal, and paediatric imaging data has gained additional interest, especially in recent years with the successful large-scale open data initiatives such as the developing Human Connectome Project, the Baby Connectome Project, and the NIH-funded Human Placenta Project. These projects enable researchers without access to perinatal scanning facilities to bring in their image analysis expertise and domain knowledge. This year’s workshop took place on September 18, 2022, as a satellite event of the 25th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2022). Two keynote speakers – Ting Xu (Child Mind Institute, New York, USA) and Gentaro Taga (University of Tokyo, Japan) – were invited for PIPPI 2022 to stimulate discussions, present recent research, and highlight future challenges in this field. Speakers working at the interface of clinical relevance and technical competence ensure close connection between technical, methodological research and clinical applications. Following our experiences from the workshops in 2020 and 2021, PIPPI 2022 made use of a hybrid setup, allowing participants to join either in person or online. The online

vi

Preface

set up of the workshops in 2020 and 2021 had several successful elements, allowing the participation of traditionally non-MICCAI attendees (clinicians, scientists) and the inclusion of a clinical keynote, and also helping to widen participation for those unable to travel. Continuing these elements improved the participation and accessibility of our workshop, whilst not adversely influencing the traditional workshop format for those attending on site. On site oral presentations were recorded live whilst virtual presentations were projected. All presented posters were displayed in joint virtual and physical settings. As part of our changes for 2022, PIPPI has enhanced the links with researchers working in Computer Assisted Intervention (CAI). This is becoming increasingly relevant to PIPPI as fetal and neonatal interventions become more complex and new surgical developments lead to highly specialized tools and advanced methods for surgical planning. This year, PIPPI saw the introduction of a new session format, the PIPPI Circle, a forum for open discussion among different communities researching early life. This session brought together scientists from clinics, industry, and academia to form a roundtable panel to discuss the most pressing challenges in fetal and paediatric imaging, future directions for research, and the clinical requirements from the user’s and patient’s perspective. PIPPI 2022 continued the support of the Fetal Tissue Annotation and segmentation (FeTA) challenge [1] and also added a new challenge, the BabySteps 2022 challenge [2]. Roxane Licandro and Jana Hutter acted as coordinators between the PIPPI workshop organizing team and the FeTA and BabySteps challenge team, respectively. Teaming up with the ISMRM Placenta & Fetus study group, and having Christopher Macgowan as both an organizer of PIPPI and the study group committee chair, has allowed PIPPI to foster more interactions between related but often separated fields, enabling researchers with joint interests in perinatal imaging but with diverse backgrounds to meet, interact, and develop new collaborations. Concrete topics of focus include motion correction and fetal cardiac imaging, topics of huge importance for the community and where excellent expertise is present within the ISMRM community. This year PIPPI teamed up with the FIT’NG (Fetal Infant Toddler Neuroimaging Group) network, an organization devoted to the study of brain development during the fetal, infant, and toddler periods. This enabled the workshop to have access to an additional community and supported the popular PIPPI topic of neuroimaging, by providing both reviewers and access to the widely distributed FIT’NG network. PIPPI 2022 received original, innovative, and mathematically rigorous papers for the analysis of both imaging data and the application of surgical and interventional techniques applied to fetal and paediatric conditions. The methods presented in these papers, and hence these proceedings, cover the full scope of medical image analysis: segmentation, registration, classification, reconstruction, atlas construction, tractography, population analysis and advanced structural, and functional and longitudinal modeling, all with an application to younger cohorts or to the long-term outcomes of perinatal conditions. All papers were reviewed by three expert reviewers

Preface

vii

from the Program Committee and ten papers were selected for presentation at PIPPI 2022 and are thus included in these proceedings. We are grateful to everyone who helped make this year’s workshop a success. September 2022

Jana Hutter Roxane Licandro Andrew Melbourne Esra Abaci Turk Christopher Macgowan

References 1. Payette, K., Steger, C., de Dumast, P., Jakab, A., Cuadra, M.B., Vasung, L., Licandro, R., Barkovich, M., Li, H.: Fetal Tissue Annotation Challenge (March 16, 2022). https://doi.org/10. 5281/zenodo.6683366. 2. Edwards, A.D., Rueckert, D., Smith, S.M., Abo Seada, S., Alansary, A., Almalbis, J., Allsop, J., Andersson, J., Arichi, T., Arulkumaran, S., Bastiani, M., Batalle, D., Baxter, L., Bozek, J., Braithwaite, E., Brandon, J., Carney, O., Chew, A., Christiaens, D., Chung, R., Colford, K., Cordero-Grande, L., Counsell, S.J., Cullen, H., Cupitt, J., Curtis, C., Davidson, A., Deprez, M., Dillon, L., Dimitrakopoulou, K., Dimitrova, R., Duff, E., Falconer, S., Farahibozorg, SR., Fitzgibbon, S.P., Gao, J., Gaspar, A., Harper, N., Harrison, S.J., Hughes, E.J., Hutter, J., Jenkinson, M., Jbabdi, S., Jones, E., Karolis, V., Kyriakopoulou, V., Lenz, G., Makropoulos, A., Malik, S., Mason, L., Mortari, F., Nosarti, C., Nunes, R.G., O’Keeffe, C., O’Muircheartaigh, J., Patel, H., Passerat-Palmbach, J., Pietsch, M., Price, A.N., Robinson, E.C., Rutherford, M.A., Schuh, A., Sotiropoulos, S., Steinweg J., Teixeira R.P.A.G., Tenev T., Tournier J-D., Tusor N., Uus A., Vecchiato K., Williams L.Z.J., Wright R., Wurie J. and Hajnal J.V.: The Developing Human Connectome Project Neonatal Data Release. Frontiers in Neurocience 16 (2022). https:// doi.org/10.3389/fnins.2022.886772.

Organization

Program Committee Chairs Jana Hutter Roxane Licandro Andrew Melbourne Esra Abaci Turk Christopher Macgowan

King’s College London, UK Medical University of Vienna, Austria, and MGH, Harvard Medical School, USA King’s College London, UK Boston Children’s Hospital, USA University of Toronto, Canada

Program Committee Alena Uus Athena Taymourtash Daniel Sobotka Dimitra Flouri Elisenda Eixarch Gemma Piella Hongwei Li Jeffrey N. Stout Kelly M. Payette Lana Vasung Lilla Zöllei Logan Z. J. Williams Malte Hoffmann Mazdak Abulnaga Pablo-Miki Martí Veronika A. Zimmer

King’s College London, UK Medical University of Vienna, Austria Medical University of Vienna, Austria King’s College London, UK Barcelona Children’s Hospital, Spain Universitat Pompeo Fabra, Spain TU Munich, Germany, and University of Zurich, Switzerland Boston Children’s Hospital, USA Children’s Hospital Zurich, Switzerland, and MGH, Harvard Medical School, USA Boston Children’s Hospital, USA MGH, Harvard Medical School, USA King’s College London, UK MGH, Harvard Medical School, USA MIT, USA Universitat Pompeo Fabra, Spain TU Munich, Germany

Contents

Automatic Segmentation of the Placenta in BOLD MRI Time Series . . . . . . . . . . S. Mazdak Abulnaga, Sean I. Young, Katherine Hobgood, Eileen Pan, Clinton J. Wang, P. Ellen Grant, Esra Abaci Turk, and Polina Golland

1

A Fast Anatomical and Quantitative MRI Fetal Exam at Low Field . . . . . . . . . . . Jordina Aviles, Kathleen Colford, Megan Hall, Massimo Marenzana, Alena Uus, Sharon Giles, Philippa Bridgen, Mary A. Rutherford, Shaihan J. Malik, Joseph V. Hajnal, Raphael Tomi-Tricot, and Jana Hutter

13

Automatic Fetal Fat Quantification from MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Netanell Avisdris, Aviad Rabinowich, Daniel Fridkin, Ayala Zilberman, Sapir Lazar, Jacky Herzlich, Zeev Hananis, Daphna Link-Sourani, Liat Ben-Sira, Liran Hiersch, Dafna Ben Bashat, and Leo Joskowicz

25

Continuous Longitudinal Fetus Brain Atlas Construction via Implicit Neural Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lixuan Chen, Jiangjie Wu, Qing Wu, Hongjiang Wei, and Yuyao Zhang

38

Automated Segmentation of Cervical Anatomy to Interrogate Preterm Birth . . . . Alicia B. Dagle, Yucheng Liu, David Crosby, Helen Feltovich, Michael House, Qi Yan, Kristin M. Myers, and Sachin Jambawalikar

48

Deep Learning Framework for Real-Time Fetal Brain Segmentation in MRI . . . Razieh Faghihpirayesh, Davood Karimi, Deniz Erdo˘gmu¸s, and Ali Gholipour

60

Attention-Driven Multi-channel Deformable Registration of Structural and Microstructural Neonatal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Irina Grigorescu, Alena Uus, Daan Christiaens, Lucilio Cordero-Grande, Jana Hutter, Dafnis Batalle, A. David Edwards, Joseph V. Hajnal, Marc Modat, and Maria Deprez Automated Multi-class Fetal Cardiac Vessel Segmentation in Aortic Arch Anomalies Using T2-Weighted 3D Fetal MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Paula Ramirez Gilliland, Alena Uus, Milou P. M. van Poppel, Irina Grigorescu, Johannes K. Steinweg, David F. A. Lloyd, Kuberan Pushparajah, Andrew P. King, and Maria Deprez

71

82

xii

Contents

Segmentation of Periventricular White Matter in Neonatal Brain MRI: Analysis of Brain Maturation in Term and Preterm Cohorts . . . . . . . . . . . . . . . . . . Alena U. Uus, Mohammad-Usamah Ayub, Abi Gartner, Vanessa Kyriakopoulou, Maximilian Pietsch, Irina Grigorescu, Daan Christiaens, Jana Hutter, Lucilio Cordero Grande, Anthony Price, Dafnis Batalle, Serena Counsell, Joseph V. Hajnal, A. David Edwards, Mary A. Rutherford, and Maria Deprez

94

Knowledge-Guided Segmentation of Isointense Infant Brain . . . . . . . . . . . . . . . . . 105 Jana Vujadinovic, Jaime Simarro Viana, Ezequiel de la Rosa, Els Ortibus, and Diana M. Sima Correction to: Knowledge-Guided Segmentation of Isointense Infant Brain . . . . Jana Vujadinovic, Jaime Simarro Viana, Ezequiel de la Rosa, Els Ortibus, and Diana M. Sima

C1

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

Automatic Segmentation of the Placenta in BOLD MRI Time Series S. Mazdak Abulnaga1(B) , Sean I. Young1,2 , Katherine Hobgood1 , Eileen Pan1 , Clinton J. Wang1 , P. Ellen Grant3 , Esra Abaci Turk3 , and Polina Golland1 1

2

Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology, Cambridge 02139, USA {abulnaga,siyoung,khobgood,eileenp}@mit.edu, {clintonw,polina}@csail.mit.edu MGH/HST Martinos Center for Biomedical Imaging, Harvard Medical School, Boston, MA 02129, USA 3 Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Harvard Medical School, Boston, MA 02115, USA {ellen.grant,esra.abaciturk}@childrens.harvard.edu

Abstract. Blood oxygen level dependent (BOLD) MRI with maternal hyperoxia can assess oxygen transport within the placenta and has emerged as a promising tool to study placental function. Measuring signal changes over time requires segmenting the placenta in each volume of the time series. Due to the large number of volumes in the BOLD time series, existing studies rely on registration to map all volumes to a manually segmented template. As the placenta can undergo large deformation due to fetal motion, maternal motion, and contractions, this approach often results in a large number of discarded volumes, where the registration approach fails. In this work, we propose a machine learning model based on a U-Net neural network architecture to automatically segment the placenta in BOLD MRI and apply it to segmenting each volume in a time series. We use a boundary-weighted loss function to accurately capture the placental shape. Our model is trained and tested on a cohort of 91 subjects containing healthy fetuses, fetuses with fetal growth restriction, and mothers with high BMI. We achieve a Dice score of 0.83 ± 0.04 when matching with ground truth labels and our model performs reliably in segmenting volumes in both normoxic and hyperoxic points in the BOLD time series. Our code and trained model are available at https:// github.com/mabulnaga/automatic-placenta-segmentation.

Keywords: Placenta

1

· Segmentation · BOLD MRI · CNN

Introduction

The placenta is an organ that provides oxygen and nutrients to support fetal growth. Placental dysfunction can cause pregnancy complications and can affect fetal development, so there is a critical need to assess placental function in vivo. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  R. Licandro et al. (Eds.): PIPPI 2022, LNCS 13575, pp. 1–12, 2022. https://doi.org/10.1007/978-3-031-17117-8_1

2

S. M. Abulnaga et al.

(a) BOLD signals increase during hyperoxia (b) Placental deformation from fetal motion

Fig. 1. Example images and placental segmentations: (a) signal brightening during hyperoxia, and (b) shape deformation caused by fetal motion. Placental boundaries are marked in yellow. Areas outside of the placenta are darkened for illustration. Intensity scale is based on the first MRI volume in the time series. (Color figure online)

Blood oxygen level dependent (BOLD) MRI can directly quantify oxygen transport within the placenta [3,16] and has emerged as a promising tool to study placental function. Temporal analysis of BOLD MRI with maternal oxygenation has been used to identify contractions [1,13], biomarkers of fetal growth restriction [7,15], predict placental age [10] and to study congenital heart disease [18,24] among many uses. Despite its importance for many downstream clinical research tasks, placental segmentation is often performed manually and can take a significant amount of time, even for a trained expert. For BOLD MRI studies, manual segmentation is rendered more challenging due to the sheer number of MRI scans acquired and rapid signal changes due to the experimental design. Experiments acquire several hundred whole-uterus MRI scans to observe signal changes in three stages: i) normoxic (baseline), ii) hyperoxic, and iii) return to normoxic. During the hyperoxic stage, the BOLD signals increase rapidly, leading to hyperintensity throughout the placenta. Furthermore, the placental shape can undergo large deformation caused by maternal breathing, contractions, and fetal motion which can be particularly increased during hyperoxia [25]. See Fig. 1 for two examples. The current practice is to analyze BOLD signals with respect to one template volume. Deformable registration of all volumes in the time series to the template is performed to enable spatiotemporal analysis [2,25]. However, due to significant motion, registration can lead to large errors, requiring outlier detection and possibly rejecting a significant number of volumes [2,25]. To address these challenges, we propose a model to automatically segment the placenta in BOLD MRI time series. Our model is trained on several volumes from each patient during the normoxic and hyperoxic phases, to capture the nuanced placental changes. We apply our model on unseen BOLD MRI volumes to demonstrate consistency in the predicted segmentation label maps. Our method performs favorably against the state-of-the-art on a large dataset with a broad range of gestational ages and pregnancy conditions. Automatic segmentation is necessary for whole-organ signal analysis, and can be used to improve time-series registration to enable localized analysis. Furthermore, it is an essential step in several post-processing tasks, including motion correction [2], reconstruction [21], and mapping to a standardized representation [4,8].

Automatic Segmentation of the Placenta in BOLD MRI

3

Machine learning segmentation models for the placenta have been previously proposed and include both semi-automatic [23] and automatic [5,10,17,19] approaches. While semi-automatic methods have achieved success in predicting segmentation label maps with high accuracy, these approaches are infeasible for segmenting BOLD MRI time series due to the large number of volumes. The majority of automatic methods focus on segmentation in anatomical images. Alansary et al. [5] proposed a model for segmenting T2-weighted (T2w) images based on a 3D CNN followed by a dense CRF for segmentation refinement and validated on a singleton cohort that included patients with fetal growth restriction (FGR). Torrents-Barrena et al. [19] proposed a model based on superresolution and an SVM and validated on a singleton and twin cohort of T2w MRI. Spektor-Fadida et al. [17] tackled the problem of domain transfer by a self-training model and demonstrated successful segmentation of FIESTA and TRUFI sequences. For a more detailed treatment of segmentation methods in fetal MRI, we refer the reader to the survey by Torrents-Barrena et al. [20]. Functional images of the placenta differ greatly from anatomical images, as they have lower in-plane resolution and the contrast between the placental boundary and surrounding anatomy is less pronounced. Anatomical images may also benefit from super-resolution approaches to increase SNR in the acquired image [21]. Pietsch et al. [10] are the first to consider placental segmentation in functional MRI. They proposed a 2D patch-based U-Net model for functional image segmentation and demonstrated a successful application of age prediction using the estimated T2* values. They focused on a cohort of singleton subjects, and demonstrated success on abnormal pregnancy conditions including preeclampsia. In contrast to their approach that segments derived T2* maps, we evaluate our segmentation model on BOLD MRI time series. Furthermore, our 3D model operates on the entire volume rather than patches, thereby helping to better resolve the boundaries of the placenta. To capture the large signal changes and placental shape variation in the time series, we train with a random sampling of manual segmentations of several volumes in the BOLD MRI series. We propose a boundary weighted loss function to more easily identify the placental boundary and improve segmentation accuracy. Finally, to evaluate the feasibility of our method for clinical research, we propose additional metrics to evaluate performance on the whole MRI time series, and illustrate a possible clinical research application.

2

Methods

We aim to find a model Fθ : X → Y that takes a BOLD MRI time series X ∈ RT ×H×W ×D and predicts a set of placenta segmentation label maps for each time point t ∈ {1, . . . T }, Y ∈ {0, 1}T ×H×W ×D , where T is the total number of time points at which MRI scans were acquired. For a given BOLD time series, we have a small number Nl of frames with ground truth labels (x, y), where x ∈ RH×W ×D is an MRI scan and y ∈ {0, 1}H×W ×D is the ground truth placenta label map.

4

S. M. Abulnaga et al.

Fig. 2. 3D Placenta Segmentation U-Net. We use a five-level 3D U-Net with maxpooling, skip connections, and convolution-transpose layers. Numbers above vertical bars denote the number of features at various stages of the processing pipeline. Batch norm is employed for normalization (batch size = 8).

2.1

Model

We use a 3D U-Net [12] with 4 blocks in the contracting and expanding paths. Each block consists of two sets of 3 × 3 × 3 convolution with ReLU activations, followed by max pooling (contraction path) or transpose convolution (expansion path), as illustrated in Fig. 2. We augment the images using random affine transforms, flips, whole-image brightness shifts, contrast changes, random noise, and elastic deformations, using TorchIO [11]. We simulate the effects of maternal normoxia and hyperoxia with a constant intensity shift in the placenta. To capture the MRI signal and placental shape changes resulting from maternal hyperoxia and fetal motion, we enhance our training with several manually segmented volumes in the normoxic or hyperoxic phase. This allows the model to learn from the realistic variations that arise during maternal oxygenation. 2.2

Additive Boundary Loss

The placental boundary can be difficult to distinguish in BOLD MRI scans due to similar appearance with surrounding anatomy. To emphasize the boundary details, we construct an additive boundary-weighting W to the segmentation loss function L. Given a ground truth placental label map y, we denote its boundary as ∂y. We use a signed distance function f (x) that measures the signed distance, d(x, ∂y), of voxel x ∈ R3 to the boundary, where f (x) < 0 when outside of the placenta and f (x) > 0 when inside. The boundary weighting is additive for voxels within δ-distance of ∂y, ⎧ ⎪ ⎨w1 if − δ < f (x) < 0, Wδ (x) = w2 if 0 ≤ f (x) < δ, (1) ⎪ ⎩ 0 otherwise. The weighted-loss is then Lw (x) = L (x) [1 + Wδ (x)] .

(2)

In practice, we set w1 > w2 , to penalize outside voxels more heavily and learn to distinguish the placenta from its surrounding anatomy. To find voxels with

Automatic Segmentation of the Placenta in BOLD MRI

5

|f (x)| < δ, we estimate a 2δ-wide boundary by an average pooling filter on y with kernel size K and take the smoothed outputs to lie in the boundary. A larger K produces a wider boundary, penalizing more misclassified voxels. 2.3

Implementation Details

We train using a learning rate η = 10−4 for 3000 epochs and select the model with the best Dice score on the validation set. For the additive boundary loss, we set w1 = 40, w2 = 1, and K = 11. All volumes are normalized by mapping the 90th percentile intensity value to 1. We use a batch size of 8 MRI volumes. We crop or pad all volumes in the dataset to have dimension 112×112×80, and train on the entire 3D volume. We augment our data with random translations of up to 10 voxels, rotations up to 22◦ , Gaussian noise sampled with μ = 0, σ = 0.25, elastic deformations with 5 control points and a maximum displacement of 10 voxels, whole volume intensity shifts up to ±25%, and whole-placenta intensity shifts of ±0.15 normalized intensity values. These values were determined by crossvalidation on the training set. When evaluating the model on our test set, we post-processed produced label maps by taking the largest connected component to eliminate islands. Our code and trained model are available at https://github. com/mabulnaga/automatic-placenta-segmentation.

3 3.1

Model Evaluation Data

Our dataset consists of BOLD MRI scans taken from two clinical research studies. Data was collected from 91 subjects of which 78 were singleton pregnancies (gestational age (GA) at MRI scan of 23wk5d – 37wk6d), and 13 were monochorionic-diamniotic (Mo-Di) twins (GA at MRI scan of 27wk5d – 34wk5d). Of these, 63 were controls, 16 had fetal growth restriction (FGR), and 12 had high BMI (BMI > 30). Obstetrical ultrasound was used to classify subjects with FGR. For singleton subjects, classification was done based on having fetuses with estimated weight less than the 10th percentile. For twin subjects, FGR classification was determined by provene monoochorionicity and discordance in the estimated fetal weight by i) growth restriction (0.9) for all but one case, with wide density at high Dice scores (>0.9. Dice differences are highly affected by fetal and maternal motion that causes placental deformation. We visually verified that modest drops in Dice (